Tools

Every tool is read-only and reads within the connection’s organization (see Connect your agent). Tools that read a workspace or project take optional workspace and project arguments; without them, they use the connection’s defaults. Your agent sees each tool’s full description and argument schema when it connects; this page is the overview.

Discovery

Tool	Returns
list_workspaces	The workspaces you’re a member of in the connection’s organization, with their project counts and your role.
list_projects	The projects in a workspace, with their run and experiment counts.
list_metrics	Metrics logged in the project: name, scale, value kind and the label keys their series are split by. The usual first call.
list_series	Concrete series (metrics with label values pinned), with the key the fetch tools return them under.
list_series_label_values	The values each label key takes in one run.
list_attribute_definitions	Attribute paths and types on runs (or the project), such as config/lr or sys/tags.
list_experiments	Experiment names.
list_runs	Runs as a table, filtered and sorted by attributes or metric summaries, with chosen attribute and summary columns. The main tool for comparing runs.
get_run_attributes	One run’s attribute values: config, status, tags, timings.

Metrics

Tool	Returns
fetch_float_series_buckets	Series compressed into buckets on a shared axis (step, wall clock or time since start), each with mean, min, max and count. The right tool for trends and whole-run curves.
fetch_float_series	Exact points, the last 200 per series by default. A step range too wide to return in full comes back bucketed instead, and says so. Histogram series can be read as a statistic (quantile, mean, …) per step.
fetch_series_summaries	Summary values (last, min, max, avg, …) per series, e.g. the final loss of every worker.
fetch_histogram_series	Whole histograms: bin edges, counts or densities, and their statistics.
fetch_scatter_series	Scatter point sets per step.

Reinforcement learning

Tool	Returns
get_rl_info	A run’s environment count and latest episode and RL step.
list_rl_environments	A run’s environments as a table, filtered and sorted like list_runs.
list_rl_metrics	Metrics logged inside environments, with their scales.
list_rl_attribute_definitions	Attribute paths logged on environments.
get_rl_environment_attributes	One environment’s attribute values, which are versioned by episode.
fetch_rl_env_series_buckets	Bucketed series for chosen environments on a shared axis.
aggregate_rl_env_series	One series reduced across environments (avg, min, max, sum, std, median) over episodes or RL steps, optionally split into groups by attribute.

RL series always need their scale: ENVIRONMENT_STEP for per-step values, EPISODE for per-episode ones.

Documentation

Tool	Returns
list_docs	Every page on this documentation site, with its title and URL.
get_doc	One page’s content as markdown.

Filters

Tools that select runs or environments take a filter: a list of conditions that must all hold. Each condition is one of:

• an attribute condition, {"attr": "config/lr", "op": "lt", "value": 0.001}, with op one of eq, ne, gt, gte, lt, lte, matches (RE2 regex), contains, contains_any, contains_all, in, not_in or exists;
• a series condition on a summary of a metric, {"series": "val/loss", "function": "last", "op": "lt", "value": 0.5}, optionally with scale and labels, or "op": "exists";
• a group {"any_of": [...]} of attribute or series conditions, at least one of which must hold. Groups don’t nest.

For example, AdamW runs that either used a small learning rate or reached a low validation loss:

[
  {"attr": "config/optimizer", "op": "eq", "value": "adamw"},
  {"any_of": [
    {"attr": "config/lr", "op": "lt", "value": 0.001},
    {"series": "val/loss", "function": "last", "op": "lt", "value": 0.5}
  ]}
]

Sorting takes {"attr": "sys/creation_time", "direction": "desc"} or {"series": "val/loss", "function": "min", "direction": "asc"}.

Responses

Responses are compact JSON. Tables come back as rows; series come back as one entry per run and series, with parallel arrays:

{"run": "run-42", "series": "train/loss", "step": [0, 100, 200], "value": [2.31, 1.07, 0.88]}

Numbers are rounded to six significant figures, and missing points are left out. Every response states how many items it holds; one cut to fit the limits is marked truncated and carries a hint. Listings of projects, runs and environments also give total_count and a next_offset for the next page.

Errors name the problem without exposing internals: an invalid argument says what was wrong with it, and a missing run or project is named.